If you can't take your preferences and history to a new AI assistant, switching gets expensive — so could a few giants end up owning the market?

This explores why AI agents might end up controlled by a few big platforms if your data, preferences and history can't move with you from one agent to another. The corpus gives a fairly direct answer. Nothing about the technology forces concentration. It happens when switching costs and control over the connecting layers pile up with whoever already owns the ecosystem. The UK's Competition and Markets Authority treats data portability as basic market plumbing, not a convenience feature Should data portability be infrastructure for agent-driven markets?. An agent is only as useful as what it knows about you. If that knowledge is stuck inside one provider, a competing agent starts out worse at helping you no matter how good it is, so leaving gets more expensive the longer you stay.

The second force is about who controls the connections between agents. Rothschild and colleagues argue that shared technical standards let agents talk to each other, but they don't decide whether the economy turns out open or closed. That depends on who controls the layer where agents connect Will agent economies develop as open webs or closed gardens?. Apple, Google, Meta and Microsoft already run the app stores, operating systems and identity systems that agents will pass through. They could rebuild app-store-style gatekeeping one level up. Portability matters here because it is what makes a standard actually usable for switching, rather than just existing on paper.

A less obvious point: big ecosystems pull agents in partly because they supply conditions that capability alone can't. One historical analysis, running from GPS to today's assistants, finds that capable agents stall when five conditions are missing: real value, personalization, trustworthiness, social acceptance and standardization Why do capable AI agents still fail in real deployments?. An incumbent can bundle all five. Personalization especially is built from the user data that portability would free up. A small agent can be technically better and still lose because it can't fill in the rest of that list. There's also a technical push the same way. Real tasks need many specialized agents working together, not one agent doing everything Do single agents always hit organizational limits?. Getting multiple agents to coordinate is hard and fragile Why do parallel subagents produce conflicting design decisions?, so a single vendor offering a ready-made set of agents that work together has a real head start.

The corpus also suggests a way out. Coordination standards tend to win adoption by wrapping and bridging existing protocols like MCP rather than trying to replace them Should coordination protocols wrap existing systems or replace them?. That lets openness grow step by step instead of requiring everyone to rewrite their systems at once. Shared experience also pays off: agents that pooled what they learned beat isolated ones by 14 to 20 points Does sharing experience across agents beat isolated evolution?. That cuts both ways. The ecosystem with the most agents learns fastest, which entrenches it, unless that learning can travel across ecosystem boundaries.

What you might not expect: the concentration risk is as much a question of governance as of markets. Some researchers argue that alignment has to expand beyond 'serve one user loyally' to inclusivity, visibility and liability Should AI agents serve one user or balance many interests?. In a world with a few dominant ecosystems, it gets unclear whose interests your agent is serving: yours or the platform's. Portability is one of the few levers that keeps that question answerable. The corpus is strong on the argument but thin on evidence. It has policy and position pieces, not measured cases of agent lock-in actually happening.